IEEE 802.11Be Multi-Link Operation: When the Best Could Be to Use Only a Single Interface

arXiv:2105.10199v1 [cs.NI] 21 May 2021 et,i snti h H,bti h A ae hr we where layer MAC the an in but also PHY, the introducing in streams, not enhanc is spatial those it Despite ments, 16 procedure. sounding to them channel upgrade implicit up to revised support Wi-Fi being to increase also bits are to 12 capabilities multiple-o (MIMO) (i.e., approaches multiple-input the modulation immediate Besides, high-order peak-throughput. as QAM symbol), 4096 per additionally new MHz, 320 a to up to the channels empowering bandwidth wider band, of GHz use 6 the the propose so, of TGbe adoption layer, do PHY complementary PH Regarding To both [3]. in layers [2]. MAC suggested and been communications have features of technical and different reliability latency the end-to-end the increase reduce performance, throughput medium and and (EHT). 802.11be (PHY) Throughput IEEE Electric physical High as known Extremely new of amendment, a (MAC) Institute control design (TG) access the and Group Task 2019, address a to may established (IEEE) in Engineers Electronic end, high- [1 demand that communications low-latency only To and not reliable (VR&AR) but that reality throughput examples augmented office straight-forward and remote virtual are instance, and to For gaming (WLANs) again. cloud networks limits area performance local their wireless applica- pushing modern are new of the tions appearance to the Nowadays, performance demands. its user adapt to innovating continuously is performance entirely assigned best simply the is interface. flow result, emptiest arriving the a new to a As when links. obtained distri are multiple they when networks across vulne neighboring highly of become activity flows the traffic to that congestion-awa reveal importa also more that results and However, the t confirm ones. on static light results outperform some policies shed obtained to The order wide in question. loads, a traffic under in and policies, evaluate scenarios of allocation we end, traffic that different To W article unresolved. the still maximize to is should interfaces radio performance flows different multiple traffic the over trans how over flow on distributed to traffic question capabilities High same the However, the the Extremely interfaces. from with data 802.11be provided points receive be IEEE and access will MLO, the stations Through of and amendment. part (EHT) be Throughput to proposed efrac vlain WLANs Evaluation, Performance h EE821b H ek ofrhrices the increase further to seeks EHT 802.11be IEEE The been has Wi-Fi ’90s, late the in back appearance its Since Terms Index Abstract EE821b ut-ikOeain hnthe When Operation: Multi-Link 802.11be IEEE et nomto n omncto Technologies Communication and Information Dept. etCudB oUeOl igeInterface Single a Only Use to Be Could Best Temliln prto MO sanwfeature new a is (MLO) operation multi-link —The WF ,IE 0.1e ut-ikOperation, Multi-link 802.11be, IEEE 7, —WiFi nvria opuFba(UPF) Fabra Pompeu Universitat laoLópez-RaventósAlvaro ´ .I I. [email protected] acln,Spain Barcelona, NTRODUCTION variety buted utput rable LAN ntly, this mit hat be re e- Y ]. , prto etr steMO olwn h oaino h T the of notation the following MLO, the as feature operation xasieyrvsd nti otx,Tb nrdcsthe introduces TGbe context, been this have In architecture MAC revised. multi-band exhaustively existing the faces, management. and desi access, architecture channel stations’ modes, and transmission APs’ of terms in ne challenges multi-l involves towards implementation shift this However, paradigm communications. a represents adoption its as a o eawy h etsolution. best to the always how be inte not on multiple over may anticipat traffic insights to the worthy distributing some is that It MLO conclusion provide them. across the traffic as the through distribute well interfaces as available framework, the advantag the coordinating show results of obtained tra different The considering policies. by balancing th question such across Therefore, tackle traffic question. we open paper, the an perfor- distribute still WLAN is properly interfaces, the multiple to maximize how to i.e., them da mance, use receive and to transmit how to However, interfaces radio multiple having channels/bands. operation frequency and multi-link different transmission in data the reception concurrent developing allow to APs capabilit on framework modern tri-band (MLO) working even, that is or fact dual, TGbe the incorporate from stations othe benefit and the transmitte To remaining between exchanges, [5]–[8]. data selected unused out is to carry to link tied receiver inter-band one are and seamless transmitted is, stations and That be dynamic result, a operation. not a preventing band, As can (MSDUs) single [4]. flow a units bands traffic data different same across service the MAC to multi- belonging limitation: the common supporting pres a designs for both However, architectures operation. band/multi-channel MAC two one. defines latter the of the analysis artic the and this on coordination focused features, multi-AP operation the th find multi-band/multi-channel under we Although label transmissions. multi-link paradigm concurrent a towards represents adoption which shift the to communications, refer We multi-link updates. of disruptive most the find can et nomto n omncto Technologies Communication and Information Dept. 1 oeal ocretoeainars utpeinter- multiple across operation concurrent a enable To Wi-Fi, for point breaking a is MLO of introduction The L losAsadsain oepottefc of fact the exploit to stations and APs allows MLO already standard 802.11 IEEE the version, current its Upon hogotti ae,w ilrfrt h multi-band/mult the to refer will we paper, this Throughout I M II. nvria opuFba(UPF) Fabra Pompeu Universitat ULTI [email protected] - acln,Spain Barcelona, IKOPERATION LINK oi Bellalta Boris : NOVERVIEW AN Gbe. i-channel nthis in rfaces our e eis le ies, ffic ink gn, ent ta. es w e e 1 r r concept of multi-link capable device (MLD), which consists (MPC) methodology. Basically, the SPC performs contention of a single device with multiple wireless PHY interfaces that on a unique channel, whereas in the MPC contention is provides a unique MAC instance to the upper layers. Such performed on all channels. Either using the SPC or the MPC implementation is achieved by dividing the MAC sub-layer method, if one channel wins contention the others are checked in two parts [9]. First, we find the unified upper MAC (U- during a PCF inter-frame space (PIFS) time to see if they MAC), which is the common part of the MAC sub-layer can be aggregated. As a result, channels in idle state are for all the interfaces. Apart from performing link agnostic aggregated and a synchronous transmission starts [15]. It is operations such as A-MSDU aggregation/de-aggregation and worth mention that AP MLDS may change its transmission sequence number assignment, the U-MAC implements a modes (e.g., asynchronous to synchronous, and vice versa) queue that buffers the traffic received from the upper-layers. at any time. Devices operating in a synchronous mode are That is, traffic awaits in the U-MAC before it is assigned referred to as constrained MLDs, or non-STR MLDs, since to a specific interface to be transmitted. Also, the U-MAC they are not allowed to transmit through an idle interface at provides some management functions such as multi-link the same time they are receiving through another. setup, association and authentication. On the other hand, there Regarding MLO feature management, the TGbe proposes is the independent low MAC (L-MAC), an individual part the multi-link setup process. Instead of designing new man- of the MAC sub-layer for each interface that performs link agement frames, TGbe agreed to reuse the current associa- specific functionalities. There, we find the link specific EDCA tion request/response frames by adding an extra multi-link queues (one for each access category) that hold the traffic element or field. That is, the setup (i.e., MLO capability ex- until its transmission. Also, procedures such as MAC header change) is performed jointly with the association mechanism. and cyclic redundancy check (CRC) creation/validation, in Then, through the multi-link element, AP MLDs and non- addition to management (e.g., beacons) and control (e.g., AP MLD negotiate and establish their subsequent operation RTS/CTS and ACKs) frame generation are implemented at scheme by exchanging their capabilities. For instance, they the L-MAC [10]. exchange information regarding the number of supported This two-tier MAC implementation enables frames to be links, their ability to perform STR, their transmission opera- simultaneously transmitted over multiple links, as the U- tion (asynchronous or synchronous), and other per-link infor- MAC performs the allocation and consolidation of MPDUs mation (e.g., frequency band, supported bandwidth, number when acting as transmitter and receiver, respectively. Also, of spatial streams, etc.) [16], [17]. To reduce overhead, it enables seamless transitions between links minimizing the the described multi-link setup process is proposed to be access latency, and addressing an efficient load balancing. performed only on a single link, which, indeed, will be the At a link level (i.e., in the L-MAC), channel access lowest in frequency due to propagation constraints. takes place. In current proposals, TGbe defines different The interface usage negotiation is performed by measuring channel access methods accordingly to two different trans- link qualities at all interfaces. That is, those receiving a mission modes: asynchronous and synchronous. First, there quality value above the clear channel assessment (CCA) is the asynchronous transmission mode. Under this opera- threshold are set as enabled, while disabled otherwise. Hence, tional mode, a MLD can transmit frames asynchronously on for each enabled interface, we will have an enabled link. It multiple links, while keeping for each one its own channel is worth mention that, as users may keep themselves mobile, access parameters (e.g., contention window (CW), arbitration any link listed as disabled may be added afterwards to the inter-frame spacing number (AIFSN), etc.). That is, each set of enabled links, by requesting a re-setup. link has its own primary channel, independent of the others. Also, this transmission mode allows the simultaneous trans- III. SYSTEMMODEL mission and reception (STR) capability over multiple links, A. Node placement enabling concurrent uplink and downlink communications. Such implementation has been proved to minimize latency We consider a set of IEEE 802.11be MLO-capable WLANs [11], while maximizing the throughput [12]. Ideally, the with N APs and M stations. Over a given area, APs are asynchronous mode should be selected as operational scheme, placed uniformly at random. However, we only accept the however, a MLD may not be STR-capable due to the in- generated scenario as valid if the inter-AP distance is higher device coexistence (IDC) interference. This issue is caused than 5 m to avoid unrealistic overlaps. Otherwise, we discard by an excessive power leakage between interfaces that do not it and generate a new one. Then, we decide the number of have sufficient frequency separation in their operating channel stations that will be associated to each AP, placing them (e.g., two channels in the 5 GHz band). As a result, IDC around it at a distance d within the interval [1-8] m and an prevents frame reception on one interface, during an ongoing angle θ between [0-2π], both selected uniformly at random. transmission on the other interface. To avoid the IDC issue, Regardless of its type, APs and stations are configured with TGbe proposes the synchronous mode, which relies on syn- 3 wireless interfaces (i.e., i1, i2 and i3), each one using 2 SU- chronized frame transmissions across the available interfaces. MIMO spatial streams. Additionally, we consider that each With that, APs or stations are prevented to perform an STR interface operates at a different frequency band. That is, for operation, avoiding IDC problems but at the cost of a lower each AP MLD interface, one channel c is selected uniformly network throughput [13], [14]. Under this operation mode, at random from the set of channels of each band, where Cb the channel access can be performed either following the is the set of available channels of band b. Thus, ci1 ∈ C2.4, single primary channel (SPC) or the multiple primary channel ci2 ∈ C5 and ci3 ∈ C6. The different set of channels are TABLE I: Evaluation setup D. Traffic generation and CSMA operation Parameter Description Only downlink traffic is considered. The traffic directed to Carrier frequency Depends on channel selection each station is modeled as an on/off Markovian model. In the C2.4 channel set 1 (20 MHz), 6 (20 MHz), 11 (20 MHz) on period, the AP receives a Constant Bit Ratio (CBR) traffic C 5 channel set 38 (40 MHz), 46 (40 MHz), 58 (80 MHz) flow of ℓ Mbps, whereas zero during the off period. Both C6 channel set 55 (80 MHz), 71 (80 MHz), 15 (160 MHz) System bandwidth 60 MHz/160 MHz/320 MHz on and off periods are exponentially distributed with mean AP/STA TX power 20/15 dBm duration Ton and Toff, respectively. Either Ton and Toff have Antenna TX/RX gain 0 dB been set to resemble a mixture of different types of traffic as CCA threshold -82 dBm AP/STA noise figure 7 dB in real scenarios. The value of the two parameters is presented Single user 2 in Table I. We refer to the traffic generated during the on spatial streams period as a traffic flow, and to ℓ as the required bandwidth. MPDU payload size 1500 bytes Path loss Same as [18] Regarding the CSMA/CA operation, it follows the abstrac- Avg. flow duration Ton = 1 s tion presented in [18], which considers the aggregate channel Avg. flow T = 3 s load observed by each AP to calculate the airtime that can be interarrival time off Min. contention allocated to each flow. Although the considered abstraction 15 window does not capture low-level details of the PHY and MAC layers Packet error rate 10% operation, it maintains the essence of the CSMA/CA: the Simulation time 120 s (1 per simulation) Number of simulations 100 (per evaluation point) ’fair’ share of the spectrum resources among contending APs and stations. In addition, this presented abstraction allows us to simulate larger networks, and larger periods of time, and detailed in Table I. Note that depending on the band, different so study the network performance at the flow-level. channel widths are considered. E. Traffic allocation policies B. MLO capabilities When a new traffic flow arrives at the AP, the traffic The stations’ set-up, as well as the establishment of the manager determines how it is distributed over the different enabled interfaces, is performed through the 2.4 GHz link. enabled interfaces of each station. The three traffic allocation Link qualities are exchanged, and therefore, interfaces that policies considered in this work are: meet the quality criteria (i.e., above CCA threshold) are set as • Single Link Less Congested Interface (SLCI): upon enabled, while disabled otherwise. In essence, this setup relies a new flow arrival, pick the less congested interface and on a simplified version of the MLO setup process described assign it to the incoming flow. in Section II. Additionally, we assume that all nodes perform • Multi Link Same Load to All interfaces (MLSA): an asynchronous channel access for each of their enabled upon a new flow arrival, distribute the incoming traffic interfaces, while their default policy is set as Multi Link Same flow equally between all the enabled interfaces of the Load to All interfaces (MLSA) (see Section III-E), which will receiving station. That is, let ℓ be the required bandwidth be our baseline evaluation policy unless stated differently. of the incoming flow, and Nint the number of enabled interfaces in the destination station. This is, traffic allo- C. Channel model and data rate selection cation per interface is given by: ℓi = ℓ/Nint, with ℓi the Path loss is characterized following the IEEE 802.11ax bandwidth allocated to interface i. enterprise model for a single floor environment. The path • Multi Link Congestion-aware Load balancing at flow loss between an station i and AP j is given by arrivals (MCAA): upon a new flow arrival, distribute the incoming traffic flow accordingly to the channel fc occupancy observed at the AP considering the enabled PL(di,j ) = 40.05 + 20 log + 20 log (min(di,j ,d )) 10 . 10 bp 2 4 interfaces at the receiving station. That is, let ρi the per- di,j centage of available (free) channel airtime at interface i. +(di,j >dbp) · 35 log10 + 7Wi,j dbp Then, the required bandwidth allocated to interface i is ρi given by ℓi∈J = ℓ , with J the set of enabled where fc is the AP’s carrier frequency in GHz, di,j is the P∀j∈J ρj distance between station i and AP j in meters, dbp is the interfaces at the target station. breaking point distance in meters, and Wi,j are the number While MLSA allocates traffic regardless of the congestion of traversed walls. We set the breaking point distance to 5 m level of each interface, the other policies take it into ac- and the number of traversed walls to 4. Note that the resulting count. When a new flow arrives, SLCI evaluates the current propagation losses are given in dB. congestion level of each interface to identify the emptiest The Modulation and Coding Scheme (MCS) used by each one, and so, it assigns the whole traffic flow to it. Then, we interface is selected according to the signal-to-noise ratio find MCAA, which tries to maximize the spectrum used by (SNR) between the AP and the stations. For instance, an balancing the traffic allocated to the different interfaces taking station’s interface i, with 2 spatial streams and a SNR into account their current occupancy. Under the MCAA, for of 11 dB, will achieve a data rate of 243.8 Mbps, using instance, a flow with a bandwidth requirement of 10 Mbps, modulation 1024-QAM with a coding rate of 5/6 and guard and capable to be distributed across three interfaces with interval of 3.2 µs in a 20 MHz channel. ρ1 = 0.2, ρ2 = 0.6, and ρ3 = 0.5, will be split as follows: Fig. 1: Policy implementation. From left to right: SLCI, MLSA and MCAA representation. Under the wireless medium representation, the gray area corresponds to the channel occupancy already seen by the MLD at the given interface, which can be from its neighbor APs, as well as its own ongoing flows.

1.54 Mbps allocated to the interface 1, 4.61 Mbps to the in a random fashion from the set of enabled interfaces. For interface 2, and 3.85 Mbps to interface 3. Although this MLO, the configuration of each interface is set as stated in approach seems very convenient, traffic may become more Section III-A, and both APs and stations use baseline MLSA vulnerable to contention with external networks, due to the policy. fact of being spread across multiple interfaces. Figure 1 shows Figure 2b shows the experienced channel occupancy by the different policies presented. APB. We observe that APB suffers from the well-known flow- in-the-middle effect in its 2.4 GHz interface, due to AP and IV. PERFORMANCE EVALUATION A APC keeping the wireless medium in busy state for most To assess the performance evaluation of the MLO, we use of their time. Consequently, traffic bandwidth requirements 2 the Neko simulation platform. in this interface are only met for low level demands in both approaches, since AP will start to drop data traffic as A. A controlled scenario B soon as traffic load increases. On the contrary, such effect First, we focus our evaluation on the comparison between is alleviated in 5 GHz and 6 GHz links as a consequence a MLO-capable deployment against a traditional multi-band of nodes’ spatial distribution, which provides all APs with single link (SL). To do so, we consider a controlled scenario two downlink contention-free links (i.e., only the 2.4GHz to pinpoint the potential benefits about the introduction of link satisfies the energy detection threshold). Although such MLO, while identifying some pitfalls without the complex in- condition may benefit the MLO capability implementation, in teractions that random deployments may originate. Figure 2a this case it is wasted by not using a congestion-aware traffic depicts a typical controlled WLAN deployment, in which allocation policy to leverage both contention-free links. Such three basic service sets (BSSs) are placed inline. Under such issue has a negative impact over APB performance, as its scenario, we assess the network behavior under two different experienced occupation is nearly the same either when using cases: i) high traffic demands, and ii) high station density both multi-band SL or MLO approaches. That is, there is deployments. For each evaluation point considered in each no difference between spreading in the same proportion the case, we simulate 100 scenarios in which stations’ positions traffic over the different links through MLO, than balancing are newly generated following the stated in Section III-A. and/or steering clients across links, like current multi-band To begin with, we characterize the network performance SL deployments do. Thus, it is easy to conclude that without for different traffic loads. In this regard, we fixed the average a proper traffic allocation policy, MLO underperforms. number of associated stations per AP to 20, while increasing Similarly, Figure 2c shows the aggregate throughput of the required bandwidth for each flow from 1 Mbps to 8 Mbps. each AP when the number of stations per BSS increases. Dif- That is, 800 total simulated scenarios, as we evaluate 8 ferently from the previous case, we set now that the required traffic load ranges for 100 different scenarios. Besides, the bandwidth for each generated flow ℓ is selected uniformly whole deployment is considered to be either multi-band SL from the 2 Mbps to 8 Mbps interval (i.e., ℓ ∼ U[2, 8]). The (i.e., 3 SL APs), or MLO-capable (i.e., 3 AP MLDs). In both other simulation parameters are kept as previously. Now, we scenarios, the channel configuration is the same for the 3 APs, introduce the SLCI congestion-aware policy presented in Sec- whereas the other simulation parameters are kept as presented tion III-E to provide a comparison between a non-congestion in Table I. It is worth mention that when using the multi-band and a congestion aware policy. From the figure, we observe SL approach, stations select the interface to be attached to how the MLSA policy struggles to cope even with 15 stations 2The Neko simulation platform can be found in GitHub at: per BSS. For instance, APB already registers throughput https://github.com/wn-upf/Neko losses around 12% for 15 stations. On the contrary, the SLCI ML (MLSA) - AP ML (MLSA) - AP ML (MLSA) - AP ML - 2.4 GHz SL - 2.4 GHz A B C ML - 5 GHz SL - 5 GHz ML (SLCI) - AP ML (SLCI) - AP ML (SLCI) - AP ML - 6 GHz SL - 6 GHz A B C 1 150

0.8 125

Offered load per BSS

0.6 100

75 0.4 Channel occupancy B 50

AP 0.2 Aggregate throughput [Mbps]

25 0 1 2 3 4 5 6 7 8 5 10 15 20 25 30 Required flow bandwidth [Mbps] Number of active STAs per BSS (a) (b) (c) Fig. 2: Controlled scenario evaluation. From left to right: a) scenario representation, b) average channel occupancy for AP B, and c) aggregate throughput performance. The high, medium and low shaded areas represent the operation range for the 6 GHz, 5 GHz and 2.4 GHz bands, respectively. assisted MLO operation supports a higher number of stations, 1 MLSA while APB only experiences a 5% throughput losses when SLCI considering 20 stations. Additionally, we observe that the 0.8 MCAA ﬂow-in-the-middle effect is reduced by the implementation

95) 0.6 of a congestion-aware policy. For instance, we observe that . 0

APB’s aggregate throughput is reduced 10% in comparison ≥ s to its neighbors when considering MLSA for 20 stations, ( 0.4 P whereas only a 4% when considering SLCI. 0.2 B. Random deployments In order to get more insights with respect to the perfor- 0 1 2 3 4 5 6 7 8 mance of the different policies presented in Section III-E, we Required flow bandwidth [Mbps] conduct a set of simulations in random generated scenarios. Similarly to the previous case, we address two different cases: Fig. 3: Prob. of average satisfaction i) high user demands, and ii) high AP density deployments. 2 TABLE II: Traffic allocation efficiency Both cases are conducted over an area of 45x45m , with nodes being placed and configured as stated in Section III. bandwidth Policy Besides, we consider that both APs and stations within the req. per flow evaluation area implement the same policy. Other simulation MLSA SLCI MCAA conditions are kept as stated in Table I. For each evaluation 2 Mbps 0.996 1.00 1.00 point considered in each case, we simulate 100 different 4 Mbps 0.925 0.989 0.985 scenarios in which APs and stations’ positions are generated 6 Mbps 0.830 0.931 0.930 as stated in Section III-A. 8 Mbps 0.750 0.833 0.842 First, we conduct the evaluation with respect to stations traffic requirements. For this purpose, we locate 10 AP MLDs within the considered area, as well as a random number of due to the fact that traffic allocation is no longer allocated stations (i.e., M ∼ U[15, 25]) for each AP MLD. Then, we blindly, but following a congestion-aware approach. In this carry out simulations by gradually increasing the required regard, we find that both SLCI and MCAA congestion-aware bandwidth ℓ of each incoming flow from 1 Mbps to 8 Mbps. policies are capable to deal with higher traffic loads without To evaluate the performance of the network we use the losing performance. For instance, for ℓ = 5 Mbps, SLCI satisfaction s metric. We define the satisfaction as the ratio and MCAA are able to provide excellent satisfaction values between the required airtime to achieve the required flow in 92% and 88% of the scenarios evaluated, respectively, bandwidth, and the actual amount that can be allocated. This whereas MLSA only manages to get good results in 5% of metric is calculated for each AP, and then, s indicates the them. Finally, notice that SLCI works better for medium load average satisfaction of the network computed as the sum of ranges, while MCAA works better for high loads as its ability the average satisfaction experienced by each AP divided by to distribute traffic across the different interfaces results in the total number of APs in the WLAN. Figure 3 shows the slightly higher satisfaction values. probability that the whole WLAN achieves an average satis- To further examine the obtained results, we show in Ta- faction s greater or equal than 95%. As expected, the baseline ble II the traffic allocation efficiency. We define this efficiency MLSA operation is outperformed by the rest of policies, as the ratio between the required bandwidth of each flow, and 1 1 Low Medium instantaneous state of the network are the best performing 0.75 0.75 ones. Additionally, we have observed that allocating an 0.5 0.5 incoming flow to the emptiest interface is almost as good, MLSA 0.25 SLCI 0.25 if not better, than proportionally distributing the flow over MCAA 0 0 multiple interfaces. Such conclusion relies on the fact that a 0 10 20 30 40 0 10 20 30 40

CDF 1 1 single interface policy not only reduces trafﬁc exposure, but Med-high High 0.75 0.75 minimizes the complexity of the trafﬁc allocation procedure.

0.5 0.5 Regarding future work, we consider three potential ex- tensions to further improve MLO operation. First, extend 0.25 0.25 MCAA to re-allocate the traffic over the different interfaces 0 0 0 10 20 30 40 0 10 20 30 40 periodically. Indeed, such implementation may minimize the Average throughput drop ratio [%] negative impact of the activity of neighboring networks in Fig. 4: Drop ratio CDF for different network densities terms of performance, but at the cost of entailing a much complex policy structure. Second, throughout this paper, we only considered IEEE 802.11be MLO capable networks, the actual throughput, for each of the policies. Once more, we whereas in real scenarios they may coexist with legacy can see that congestion-aware policies are able to outperform single link networks. Therefore, it may be relevant to study the baseline operation, as all of them can effectively allocate the different MLO traffic allocation policies under those more than the 90% of the required bandwidth per flow for conditions. Third, and last, to integrate the MLO operation in values between 2 Mbps and 6 Mbps. Also, notice how the a Software Defined Networking framework [19] to orchestrate performance of the two congestion-aware policies (i.e., SLCI the different APs that belong to the same WLAN, as well as and MCAA) is highly similar. to design and evaluate centralized multi-AP MLO policies. Finally, we assess the performance of the different policies under high AP density deployments. To do so, we gradually ACKNOWLEDGMENTS increase the number of APs, to generate four different density This work has been partially supported by the Spanish deployments: low (N = 5), medium (N = 10), med-high Government under grant WINDMAL PGC2018-099959-B- (N = 20) and high (N = 40). For each AP, we placed a I00 (MCIU/AEI/FEDER,UE), by the Catalan Government random number of stations (i.e., M ∼ U[15, 25]). Figure 4 under grant 2017-SGR-1188, and by a Gift from Cisco shows the empirical cumulative distribution function (CDF) (2020). of the average throughput drop ratio experienced for each density. The throughput drop ratio represents the percentage REFERENCES of traffic that cannot be served. Then, it is computed as one [1] M. Carrascosa and B. Bellalta, “Cloud-gaming: Analysis of Google minus the ratio between the achieved throughput and the re- Stadia traffic,” arXiv preprint arXiv:2009.09786, 2020. quired one. As expected, the probability of having higher drop [2] T. Adame, M. Carrascosa, and B. Bellalta, “Time-sensitive networking in IEEE 802.11 be: On the way to low-latency WiFi 7,” arXiv preprint ratio values increases upon network densification. However, arXiv:1912.06086, 2019. the different congestion-aware policies are able to reduce [3] D. López-Pérez, A. Garcia-Rodriguez, L. Galati-Giordano, M. Kasslin, significantly those values. For instance, for medium density and K. Doppler, “Ieee 802.11 be extremely high throughput: The next generation of wi-fi technology beyond 802.11 ax,” IEEE Communica- deployments, we find that drop ratio is decreased 2.25x tions Magazine, vol. 57, no. 9, pp. 113–119, 2019. by means of SLCI implementation for the 75th percentile [4] Y. Fang, B. Sun, H. Zhiqiang, and N. Li, “Multi- compared to MLSA. Additionally, we find that SLCI further link architecture and requirement discussion.” IEEE 802.11 Documents. 2019. [Online]. Available: improves MCAA as the density increases. This phenomena https://mentor.ieee.org/802.11/documents?is dcn=1095&is group=00be. occurs as MCAA policy allocates traffic to more interfaces, [5] Z. M. Fadlullah, Y. Kawamoto, H. Nishiyama, N. Kato, N. Egashira, which in fact increases the exposure to the dynamics of the K. Yano, and T. Kumagai, “Multi-hop wireless transmission in multi- band wlan systems: Proposal and future perspective,” IEEE Wireless other APs. That is, in high density deployments, APs tend Communications, vol. 26, no. 1, pp. 108–113, 2017. to have more neighboring APs, and so, the use of a high [6] P. Kyasanur and N. H. Vaidya, “Routing and interface assignment number of interfaces to transmit makes the traffic much more in multi-channel multi-interface wireless networks,” in IEEE Wireless Communications and Networking Conference, 2005, vol. 4, pp. 2051– vulnerable as the probability of having a channel overlap 2056, IEEE, 2005. increases. [7] H. Singh, J. Hsu, L. Verma, S. S. Lee, and C. Ngo, “Green operation of multi-band wireless lan in 60 ghz and 2.4/5 ghz,” in 2011 IEEE Con- sumer Communications and Networking Conference (CCNC), pp. 787– V. CONCLUSIONS & FUTURE WORK 792, IEEE, 2011. In this paper, we assessed and evaluated the adoption of [8] N. Choi, Y. Seok, and Y. Choi, “Multi-channel mac protocol for mobile ad hoc networks,” in 2003 IEEE 58th Vehicular Technology Conference. multi-link communications by the upcoming IEEE 802.11be VTC 2003-Fall (IEEE Cat. No. 03CH37484), vol. 2, pp. 1379–1382, amendment. Through a wide variety of scenarios, our results IEEE, 2003. show that the implementation of MLO can help the next [9] Y. Fang, B. Sun, H. Zhiqiang, and L. Lu, “Multi-link reference model discussion.” IEEE 802.11 Documents. 2020. [Online]. Available: generation Wi-Fi networks to satisfy the highly demanding https://mentor.ieee.org/802.11/documents?is dcn=1131&is group=00be&is year=2020. requirements of modern applications. However, we show that [10] J. Levy and X. Wang, “802.11be architecture/association the performance of MLO depends mainly on the imple- discussion.” IEEE 802.11 Documents. 2020. [Online]. Available: https://mentor.ieee.org/802.11/documents?is dcn=1122&is group=00be. mented traffic allocation policy. From that side, as expected, [11] G. Naik, D. Ogbe, and J.-M. J. Park, “Can wi-fi 7 support real-time congestion-aware allocation policies able to adapt to the applications? on the impact of multi link aggregation on latency,” [12] S. Naribole, W. B. Lee, S. Kandala, and A. Ranganath, “Simultaneous [16] I. Jang, J. Choi, J. Kim, S. Kim, S. Park, and T. Song, “Discussion on transmit-receive multi-channel operation in next generation wlans,” multi-link setup.” IEEE 802.11 Documents. 2019. [Online]. Available: in 2020 IEEE Wireless Communications and Networking Conference https://mentor.ieee.org/802.11/documents?is dcn=1509&is group=00be. (WCNC), pp. 1–8, IEEE, 2020. [17] I. Jang, J. Choi, J. Kim, T. Song, and N. Kim, [13] W. Murti, S. Kim, J.-H. Yun, J. Son, G. Ko, and “Indication of multi-link information: Follow-up.” K. Jin Sam, “Performance and fairness of multi-link IEEE 802.11 Documents. 2020. [Online]. Available: operations.” IEEE 802.11 Documents. 2019. [Online]. Available: https://mentor.ieee.org/802.11/documents?is dcn=741&is group=00be. https://mentor.ieee.org/802.11/documents?is dcn=1633&is group=00be. [18] A.´ López-Raventós and B. Bellalta, “Concurrent decentralized channel [14] A. Patil, G. Cherian, A. Asterjadhi, and D. Ho, “Multi-link aggregation allocation and access point selection using multi-armed bandits in multi - gain analysis.” IEEE 802.11 Documents. 2019. [Online]. Available: bss wlans,” Computer Networks, vol. 180, p. 107381, 2020. https://mentor.ieee.org/802.11/documents?is dcn=764&is group=00be&is [19]year=2019.A.´ López-Raventós, F. Wilhelmi, S. Barrachina-Muñoz, and B. Bellalta, [15] I. Jang, J. Choi, J. Kim, S. Kim, S. Park, and “Combining software defined networks and machine learning to enable T. Song, “Channel access for multi-link operation.” self organizing WLANs,” in 2019 International Conference on Wireless IEEE 802.11 Documents. 2019. [Online]. Available: and Mobile Computing, Networking and Communications (WiMob), https://mentor.ieee.org/802.11/documents?is dcn=1144&is group=00be. pp. 1–8, IEEE, 2019.